Several ways to act on it.
| 01 | Letter — the visibility you can’t see is the kind that hurts |
| 02 | The main report — what ‘cited’ actually means, and how often it’s wrong |
| 03 | Questions a skeptical CEO would raise |
| 04 | From the source file — how to read an accuracy number honestly |
| 05 | The RAG clinic — one retrieval failure, traced from symptom to repair |
| 06 | Company case file — Wolf River Electric v. Google |
| 07 | AI visibility in action — the missing / mentioned / cited / correct framework |
| 08 | This week’s action plan — five steps, one of them free |
| 09 | Services overview |
| 10 | Sources and supporting documentation |
According to the company’s complaint, in the autumn of 2024 sales representatives at a solar and electrical contractor in Isanti, Minnesota started noticing something they had no name for. Customers were backing out — in some cases cancelling signed contracts, in others refusing to proceed with proposals — late in the process, after the work had been quoted. When reps asked why, several said the same thing: they had searched the company’s name on Google, and Google had told them the firm was being sued by the Minnesota Attorney General for deceptive sales practices, hidden fees, and high-pressure tactics.
There was no such lawsuit. The company, Wolf River Electric, had never been named by the Attorney General. The actual case the AI was confusing it with had named four solar lending firms, none of them Wolf River. But Google’s AI Overview stated the claim plainly, in a calm paragraph, with sources listed underneath. The buyers read the paragraph, believed it, and walked. The reps logged the losses and moved to the next lead. Nobody on the selling side could see the cause, because the cause was a sentence written by a machine inside a search result the company never visited.
That is the problem this issue is built around, and it is not a solar problem. It is the problem every B2B vendor now has and most cannot measure: an AI engine can name you, describe you, attach a citation to the description, and be wrong — and the buyer who reads it will never tell you it happened. There is no bounced email, no abandoned cart, no line in the CRM. The deal simply does not arrive.
The comfortable belief in this market is that any mention is a good mention. Get into the answer, the thinking goes, and you have won the visibility game. I want to spend this issue taking that belief apart, because the evidence says the opposite often enough to matter to your forecast. Being absent from an AI answer costs you a chance. Being present and misdescribed costs you the buyers who were furthest along — the ones who trusted the answer enough to act on it.
I will show you what the independent research says about how often AI engines misstate their own sources, walk through the Wolf River case in the detail the court filing allows, and give you a way to measure your own exposure that does not depend on trusting a vendor’s dashboard — including ours. The aim is not to alarm you about AI. It is to replace a feel-good metric with one that tells you the truth.
— Elizabeta Kuzevska, Co-Founder, Revenue Experts AI
Start with what a citation is supposed to be. When an AI engine answers a buyer’s question, it does two separate things in the same breath. It makes a claim — “this vendor tracks brand mentions across ChatGPT and Perplexity” — and it attaches a source, a link presented as the proof of that claim. The implicit promise is: don’t just trust me, here is the page that backs this up.
The question almost no measurement tool asks is the obvious one. Does the page actually back it up? When researchers have checked, the answer is unsettling. A 2025 study published in Nature Communications built an automated pipeline called SourceCheckup and ran seven popular models across 800 medical questions and roughly 58,000 statement-source pairs. Between 50% and 90% of the responses were not fully supported, and sometimes were contradicted, by the sources the models themselves had cited. Even GPT-4o with web search left about 30% of individual statements unsupported, and nearly half of its full responses only partly supported.
One scope note, because it is the kind of thing this publication refuses to paper over: that study measured medical queries, where the standard for ‘supported’ is strict and the source literature is dense. It is not a clean read on B2B software questions. I cite it because it is peer-reviewed and it measures the exact mechanism — claim-to-source fidelity — that matters here, not because the percentage transfers unchanged to your category. Treat it as evidence that the mechanism fails routinely, not as your specific failure rate.
The failure has two layers, and they are worth separating. The first is the page that exists but does not support the claim: the AI cites a real, related page that simply does not contain the thing the AI said. The second is worse and rarer — the page that does not exist at all. Work posted to arXiv in 2026 measuring citation URLs from commercial models and ‘deep research’ agents found 3–13% of cited URLs were fabricated, with no record in the Wayback Machine that they ever existed, and 5–18% non-resolving overall. The agents that produce the most citations hallucinated URLs at the highest rates. More confident output, more sources listed, more chances to be wrong.
Now connect that to the buyer. G2 surveyed 1,076 B2B software buyers and decision-makers in March 2026 and published the findings in April 2026 — the people this report is written for — and reported that 64% encounter inaccurate AI chatbot recommendations often or very often. Not once. Routinely. The same buyers told G2 that citations from software review sites were the single most confidence-inspiring signal in an AI answer, and that when an AI conflicted with a brand they already trusted, roughly a quarter went to peer reviews to break the tie. (G2 sells into this market, so read its numbers as the survey of an interested party. I use them for the buyer-behavior direction, which independent sources corroborate, not as the last word; the 64% figure in particular I could not confirm against G2’s primary data table this session, only against secondary write-ups of the release.)
Put the two halves together, but keep them as two findings rather than one causal chain. In the medical-query study, claims that were unsupported or only partly supported by their cited sources occurred at rates measured in the tens of percent; the exact rate for B2B software questions is not established by that study. Separately, the G2 survey shows that buyers report meeting inaccurate AI recommendations often and that AI now shapes their decisions. Neither finding alone proves that a specific buyer walked because of a specific false claim — but together they describe a market where the mechanism is common and the buyers are acting inside it. The vendor who is absent from the answer loses a shot at a buyer who might not have been ready anyway. The vendor who is present and wrong risks losing the buyer who was ready — who trusted the answer enough to move.
This is why ‘any mention is a good mention’ is not a harmless simplification. It tells you to chase appearance rate and stop there. Appearance rate counts the room you are in. It does not check what was said about you in that room. A vendor optimizing purely for presence can raise the number on the dashboard while the sentences attached to that presence quietly describe the wrong product, the wrong price, or a lawsuit that was never filed.
The AI Visibility Audit runs 50 buyer questions across three engines, three times each, then opens every cited page and checks whether it supports the claim. Presence and accuracy, reported separately.
Book the audit →Most AI-visibility tools count. They run your queries, look for your name or your domain in the output, and report how many times it appeared and where it ranked. That is genuinely useful for one thing: telling you whether you are in the conversation at all. It is silent on the question that decides whether the conversation helps or hurts you — whether what was said is true.
The reason counting dominates is that it is cheap and checking is not. To check, you have to take every citation the engine produced, open the page it points to, read it, and decide whether that page actually supports the specific sentence it was attached to. Across a real run — dozens of questions, multiple engines, several repeats — that can mean reviewing hundreds of answers and potentially hundreds or thousands of citation instances, depending on how many sources each engine returns. So most tools skip it and present a count as if it were a verdict. The count says “you appeared 40 times.” It does not say whether those 40 appearances described you correctly, and it cannot, because it never opened the pages.
There is a second trap inside the counting approach: non-determinism. Ask the same engine the same question three times and you will often get three different answers, with different sources. A tool that asks once and reports the result is reporting a single roll of the dice as if it were the house edge. Repeated runs of the same buyer question can return different recommendation sets and different sources; identical-prompt reruns are not guaranteed to agree, and paraphrased versions of the same question diverge further. A measurement taken once is not measurement; it is an anecdote with a timestamp.
The honest alternative is slower and less flattering. You repeat each question enough times to separate signal from noise. You open every cited page and score whether it supports its claim. You report presence and accuracy as two different numbers, because they are two different things, and you keep them per engine, because the engines behave differently enough that blending them into one figure hides the truth rather than revealing it. The source file section that follows shows exactly how to read such numbers without fooling yourself.
The belief survives only if accuracy is free — if being named always means being named correctly. The evidence says accuracy is not free and misnaming is common. So the belief has to be replaced, not softened. The goal is not maximum appearance. It is the right appearance, with the right description, on the questions your buyers actually ask. A brand cited on every query and described as overpriced and unreliable is losing faster than a brand cited on a third of them and described correctly.
Legally, that is unsettled and being fought over right now. Commercially, it is already yours. The buyer who reads a false claim about you and cancels does not file a brief; they just leave. Whatever a court eventually decides about Google’s liability, Wolf River’s lost contracts were lost in real time, to a company that did nothing wrong. The exposure lands on you first and is litigated later, if ever.
Some do, and that is why review-site citations and peer checks matter so much in the G2 data. But ‘distrust’ here means ‘verify before fully believing,’ not ‘ignore.’ Buyers still start in the engine, still form a shortlist there, and still carry the first impression into the verification step. If the first impression is wrong about you, you are now arguing uphill against a machine the buyer half-trusts.
Partly, but the relationship depends on the engine and on how overlap is measured. Studies consistently find some connection between traditional rankings and AI citations, yet they report very different figures because they measure different systems and different denominators. Ranking well can improve your chances of being sourced; it does not guarantee inclusion, and it says nothing about whether the AI describes you correctly. You can rank first and still be described wrongly inside the AI summary that sits above your link.
Two things, in order. First, find out where you are currently misdescribed or absent on the questions that matter — you cannot fix what you have not measured. Second, strengthen the owned and third-party pages the engines lean on, so that when they reach for a source about you, the page they find says the true thing clearly. The action plan later in this issue is the tool-free version of that work.
If you are going to act on an AI-visibility number, you need to know what it does and does not say. Two distinctions do most of the work, and both are easy to get wrong.
When you check whether a cited page supports its claim, you get one of a few verdicts: it fully supports, it partly supports, it does not support, or you could not read it. Full-support rate counts only the citations whose pages fully support the attached claim. Overall support rate counts citations that fully or partly support it. The two can diverge sharply. A run might show one engine at 33% full-support but 74% overall support. That gap does not, by itself, tell you why the citations only partly support the answer. It may be a broad summary sentence with several sources hung off it; it may be weak citation placement, omitted qualifiers, or claims that reach past the evidence. The underlying examples have to be read before the gap can be interpreted — a low full-support rate is a flag to investigate, not a verdict on its own.
This is the entire reason a serious report keeps the engines separate and never blends them into one tidy accuracy figure. The engines bind citations differently. One ties each citation to a specific sentence, so its cited pages tend to match tightly. Another attaches many sources to a loose summary. Average them and you erase the real difference and report something that is true of no engine in particular.
Some cited pages cannot be read by an automated checker: they are login-walled, blocked to machines, or have been taken down. The honest move is to mark those as not-checkable and leave them out of the accuracy calculation entirely, so the percentage is computed only on pages a human or a checker actually read. The dishonest move — and it produces a prettier number — is to guess that the blocked pages probably support their claims and fold the guess into the score. When you read any vendor’s accuracy figure, ask which of those two they did. If they cannot tell you, the number is decoration.
One more caution, aimed at our own house as much as anyone’s. We currently exclude one major engine, Gemini, from our own audits. Google’s terms for Grounding with Google Search place limits on how grounded results and related search data may be used, displayed, and retained, and our current audit workflow is not built to operate within those limits. That is an operational and compliance decision, not a claim that Gemini cannot be measured by any method. We disclose it rather than bury it — and the Wolf River case happens to involve the very engine we do not measure. An audit that covers three engines should say three, not imply four.
An engine answers a category question and names your company, but the sentence is subtly wrong — it attributes a competitor’s feature to you, or states a price you retired a year ago. The citation underneath points to a real page on your own site.
This is the retrieval-augmented generation failure mode that does the most quiet damage. The engine retrieved a page that is genuinely about you, then generated a sentence that the page does not actually support, blending it with information pulled from elsewhere in its context. The page is real and relevant, so a counting tool scores it as a clean citation. The claim is false, so the buyer who reads it is misinformed. Retrieval succeeded; grounding failed. The two are not the same step, and tools that only check ‘did a relevant page get retrieved’ will miss this every time.
A common root is stale or ambiguous owned content. If your pricing page still shows last year’s number in a comparison table, or your feature list is written loosely enough that a model can mis-assign a capability, you have handed the engine the raw material for a confident error — and your own page is the citation that makes it look trustworthy.
Find the specific pages the engines reach for when they describe you — the audit’s page-concentration view shows that a small number of your pages usually carry most of your citations. Then make those pages state the load-bearing facts in plain, declarative sentences: current price, named clearly; each feature attributed to the right product; no outdated comparison left lying around for a model to misread. You are not optimizing for a keyword. You are removing the ambiguity that lets a model generate a wrong sentence and cite you as the proof of it.
A house rule of this report is that the case file must be a real, named, third-party company with a publicly documented court record — never a composite, and never our own audit dressed up as proof. Wolf River qualifies on every count. It is a roughly 200-employee Minnesota solar and electrical contractor. The facts below come from its filed complaint and from independent legal and news coverage, not from any vendor’s marketing. And one thing must stay in front of every sentence that follows: these are allegations in a pending case. Google denies them. No court has ruled on the merits. I am reporting what was filed and documented, not a finding of liability.
According to the complaint, when users searched the company’s name, Google’s AI Overview stated that Wolf River was facing a lawsuit from the Minnesota Attorney General for deceptive sales practices — misleading customers about cost savings, using high-pressure tactics, and tricking homeowners into contracts with hidden fees. No such lawsuit existed. The Attorney General’s actual March 2024 action had named four solar lending companies — GoodLeap, Sunlight Financial, Solar Mosaic, and Dividend Solar Finance. Wolf River was not among them.
That single line is why this case belongs in an issue about citation accuracy rather than one about hallucination in general. This was not an unsourced guess. The AI listed sources. The sources were real pages. The pages did not say the thing the AI attributed to them. The citation furniture — the links, the calm authoritative tone — made a fabricated claim look verified. Independent reporting noted the same: the cited materials mentioned solar-industry lawsuits, but Wolf River was not a defendant in them.
The complaint lists specific lost business, the kind of detail that turns an abstract risk into a number a CFO recognizes. A customer terminated on March 3, 2025, contract price $39,680. Another refused to proceed on March 4, a $26,400 proposal. On March 5, a customer sent a sales rep a screenshot of Google’s false statement and cancelled a $150,000 contract even after the CEO told them directly it was untrue. On March 11, a nonprofit “pulled the plug” on a $147,400 solar project plus a $26,644 lighting project, citing “several lawsuits in the last year” with the Attorney General’s office. The company’s claimed damages escalated from a $24.7 million figure cited in 2024 to an estimated range of $110 million to $210 million in mid-2025 disclosures.
In at least one of those cancellations, the complaint says the CEO personally told the customer the claim was false — and the customer left anyway. That is the part that should worry any vendor. According to the complaint, the customer cancelled even after the CEO denied the claim, suggesting that the AI-generated statement carried more weight than the company’s own correction. If your defense against an AI error is ‘we’ll just tell them it’s wrong,’ this case is a warning that the defense can fail.
It does not prove that this happens to most companies, or that any given vendor is being misdescribed today. It is one documented case, still in litigation. What it documents is a concrete allegation and a plausible mechanism: an AI system can present a false claim about a named company, attach real-looking sources, and influence a buyer before the company has a chance to respond. The case does not establish how often this happens, and no court has yet ruled that Google caused the claimed losses. Even so, a documented allegation of this kind is enough to justify measuring your own exposure rather than assuming you have none.
A readiness check shows whether your owned pages give the engines clean, current facts to cite — or ambiguity they can misread.
Run the readiness check →Most visibility dashboards stop at the first three rungs below. The argument of this issue is that the fourth rung — correctness — is the one that protects revenue, and that reporting the first three without it gives a false sense of safety. Here is the full ladder, and what to record at each step.
| 1 | Missing | The engine answers the buyer’s question and does not mention you at all. A lost chance, but a clean one — no false impression is formed. |
| 2 | Mentioned | Your name appears, with no link. You are in the conversation; you cannot yet tell whether the engine got you right. |
| 3 | Cited | Your name appears with a source attached. The dashboard usually stops here and calls it a win. |
| 4 | Correct | The cited page actually supports what the engine said about you. This is the rung that decides whether the citation helps or hurts — and the one counting tools never check. |
| 5 | Recommended | The engine not only describes you correctly but puts you on the shortlist. Only meaningful once Correct is true; a wrong recommendation is a liability, not a win. |
Record each question at its highest rung, and keep a separate column for accuracy so a CITED result that fails the CORRECT test is visibly flagged rather than silently counted as a win. A report built this way will sometimes show a lower headline number than a counting tool — because it refuses to count a wrong citation as a good one. That lower number is the honest one.
Open ChatGPT, Claude, and Perplexity. Ask each the three questions your best buyers ask before they reach you — the ‘best tool for X,’ the ‘is [your company] any good,’ the head-to-head against your main rival. Ask each one three times. Read what is said about you and, critically, open the sources and check whether they support it. This costs nothing and takes an afternoon, and it is the single most clarifying thing on this list.
For each answer, record two things, not one: were you present, and was what was said correct. Resist the urge to collapse them. The gap between ‘mentioned’ and ‘mentioned correctly’ is the whole subject of this issue.
Most of your AI visibility rides on a handful of your own pages. Find them and make them state the load-bearing facts — current price, correct feature attribution, no stale comparison tables — in plain declarative sentences a model cannot misread.
AI engines often rely heavily on pages you do not own when they describe your brand — review sites, directories, editorial pages, forums, reference sites. Review the third-party pages that repeatedly appear in your own results, and correct the ones that are wrong or outdated, instead of assuming your website controls the full description. You do not own them, but you can usually request fixes.
Because the engines are non-deterministic and change over time, a single check is a snapshot, not a trend. Put a recurring date on the calendar — monthly is enough for most — and watch the direction, not any single reading.
The full audit runs 50 questions across three engines, three times each, and opens every cited page to score support — the version of step one that becomes difficult and time-consuming to perform manually at scale.
Book the audit →Case LTL LED, LLC d/b/a Wolf River Electric v. Google LLC, Case No. 25-cv-02394 (D. Minn.); complaint quoted in The Volokh Conspiracy (Reason), June 11, 2025: reason.com/volokh/2025/06/11/large-libel-models-small-business-sues-google-claiming-ai-overview-in-searches-hallucinated-attorney-general-lawsuit/
Case Case status and damages range corroborated by Quinn Emanuel client alert, ‘Defamation in the AI Era,’ March 2026, and Star Tribune / GovTech reporting, June 2025. Remand to state court: Reason/Volokh, January 12, 2026. (Allegations in a pending case; Google denies them; no merits ruling.)
Context Minnesota Attorney General’s actual March 2024 action named GoodLeap, Sunlight Financial, Solar Mosaic, and Dividend Solar Finance — not Wolf River.
Accuracy SourceCheckup, Nature Communications, 2025: 50–90% of LLM responses not fully supported by cited sources; ~30% of GPT-4o-with-search statements unsupported. Scope: medical queries — evidence of the mechanism, not a B2B failure rate. nature.com/articles/s41467-025-58551-6
Fabrication ‘Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents,’ arXiv preprint, 2026: 3–13% of citation URLs fabricated; 5–18% non-resolving. Preprint, not yet peer-reviewed. arxiv.org/abs/2604.03173
Buyers G2 surveyed 1,076 B2B software buyers in March 2026 and published the findings in April 2026 (‘In the Answer Economy, Don’t Win the Click’): 64% encounter inaccurate AI recommendations often/very often; review-site citations the top confidence signal. Reported via the G2 release and secondary write-ups; the 64% figure was not located in G2’s primary data table in clean text this session and should be confirmed against the primary report page before it is treated as fully verified.
Consistency On run-to-run variation: repeated identical and paraphrased prompts can return different recommendation sets and sources across AI systems (2026 reproducibility research). Treated here qualitatively; no single same-answer percentage is asserted, as no primary figure was verified in clean text this session.
Self-audit AI Citation Audit, Revenue Experts AI (self-audit), June 2026. Used in this issue only to illustrate method (full-support vs. overall support rate; per-engine reporting; the ‘couldn’t check’ exclusion) and the audit’s specifications — never as third-party proof. Self-interested by definition; discounted as such.
The argument of this issue is narrow on purpose: being named by an AI engine is not the same as being named correctly, and the second is the one that protects your pipeline. You do not have to take that on faith. The afternoon check in section 08 costs nothing and tells you where you stand. Run it before you assume your visibility is working for you.
Reply to this issue if a claim here doesn’t hold — corrections are read and run down. And if a colleague who signs the contracts should see it, forward it.